ProtoMAP: prototypical network based few-shot learning for missed abortion prediction.

Abstract

Missed abortion is a prevalent issue in clinical practice, posing both physical risks to the mother and substantial psychological impact. Accurately predicting the risk of missed abortion is essential for guiding timely clinical interventions and safeguarding maternal health. Data on missed abortion are scarce and imbalanced. Given the limited clinical data and the nonlinear interrelationships among multiple features, traditional machine learning methods often fail to capture essential patterns, thereby their prediction performance is suboptimal. This paper proposes a prototype network based on few-shot learning, namely ProtoMAP. The goal is to train a missed abortion prediction model using a limited number of samples, while achieving performance comparable to models trained on large-scale datasets. Unlike previous studies, this work is the first to explore the problem of missed abortion prediction using a few-shot learning approach. A series of experiments were conducted, and the results demonstrate that the proposed ProtoMAP model significantly outperforms a range of baseline models in the task of missed abortion prediction. These results demonstrate that ProtoMAP not only supports missed abortion prediction in a few-shot learning setting, but also achieves performance that rivals or exceeds that of baseline models trained with overall data. And it demonstrates the practical utility of ProtoMAP for clinical missed abortion prediction, particularly in scenarios where data is scarce.